A general comment: I think these checks should happen inside _weights_ar_valid - this will ensure they are used for percentile as well.

Another general comment: as written, the common case has to go through a lot of checks. I think it would be better to optimize for the common case, and not worry too much about distinguishing failure cases. E.g., you can do just one evaluation with:

if not np.all(np.isfinite(weights)):
    raise ValueError("weights must be finite.")

Note that this loop can still give unexpected errors, because you are here counting on object arrays to be turned into their values as scalars. E.g.,

np.isnan(np.array([1.,None,np.inf])[1])
# TypeError: ufunc 'isnan' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

This will be an uninformative error!

I think we have two choices: just not check for object dtype, or convert to float before checking (and then passing on it that conversion fails).

Here again we could ensure the common case remains fast by doing:

if np.any(weights <= 0):
    raise ValueError("weights must be non-negative and cannot be all zero.")
    # or, more explicit error messages,
    if np.all(weights == 0):
        raise ValueError("At least one weight must be non-zero.")
    else:
        raise ValueError("Weights must be non-negative.")

Trying to keep this inline:

The issue with this is that some of the weights might be 0, but none of them are negative. So it would raise an error even though it shouldn't

You're right, I was too sloppy in writing this, the else should be elif np.any(weights <0) so that the case of some weights 0 falls through (slowly, but better than making all cases slow!).

p.s. Given this, I'd probably swap the order, i.e.,

if np.any(weights <= 0):
    # Do these checks guarded by the above `if` to avoid slowing down the common case.
    if np.any(weights < 0):
        raise ValueError("Weights must be non-negative.")
    elif np.all(weights == 0):
        raise ValueError("At least one weight must be non-zero.")

@Tontonio3 I don't see how you responded to this suggestion. Please make sure all reviewer feedback is addressed before requesting re-review.

@ngoldbaum Your're right, I forgot to implement this

I'd parametrize over np.quantile and np.percentile as well - they should have the same errors.

Also, if you pass in a list rather than an array, you could parametrize over dtype=float and dtype=object, to make this a little more readable.

Done